I think the point is that if R1 isn't possible without access to OpenAI (at low, subsidized costs) then this isn't really a breakthrough as much as a hack to clone an existing model.
R1 is--as far as we know from good ol' ClosedAI--far more efficient. Even if it were a "clone", A) that would be a terribly impressive achievement on its own that Anthropic and Google would be mighty jealous of, and B) it's at the very least a distillation of O1's reasoning capabilities into a more svelte form.
The training techniques are a breakthrough no matter what data is used. It's not up for debate, it's an empirical question with a concrete answer. They can and did train orders of magnitude faster.
Not arguing with your point about training efficiency, but the degree to which R1 is a technical breakthrough changes if they were calling an outside API to get the answers, doesn't it?
It seems like the difference between someone doing a better writeup of (say) Wiles's proof vs. proving Fermat's Last Theorem independently.
